advances in real-time rendering in games. pre-integrated skin shading eric penner

Advances in Real-Time Rendering in Games


Pre-integrated Skin ShadingEric Penner


Overview

Intro to SSS

TSD Approaches to SSS

Our Approach - Pre-Integrated Skin Shading

Extra time? Two Tricks for character shadows Masked Variance Shadow Maps

PCF2 using Pixel Quad Messages (16x depth samples)


Subsurface Scattering - TSD

Texture Space Diffusion (Borshukov et al.)




Gaussian Blur / Wrap lighting(Gosselin/Green)




Gaussian Blur / Wrap lighting(Gosselin/Green)

Efficient Rendering of Human Skin (d’Eon et al.)



NVIDIA Human Head Gold standard for quality in real-time

Helps to have high quality head scan



GPU Gems 3 - Diffusion Profile



NVIDIA Human Head Gold standard for quality in real-time

Lots of 10242 or 20482 maps

Many Gaussian blur passes

Take Home Messages Linear Lighting! Needs Good Spec! (multiple lobes)

Skin diffusion is not Gaussian (two pass blur isn’t enough)


Fast Subsurface Scattering


ShaderX7 (Hable et al.)

Make a custom jittered kernel

Unique weights for r/g/bfloat3 blurJitteredWeights[13] = { { 0.220441, 0.437000, 0.635000 }, { 0.076356, 0.064487, 0.039097 }, { 0.116515, 0.103222, 0.064912 }, { 0.064844, 0.086388, 0.062272 }, ....};

float2 blurJitteredSamples[13] = { { 0.000000, 0.000000 }, { 1.633992, 0.036795 }, { 0.177801, 1.717593 }, { -0.194906, 0.091094 }, ....};




ShaderX7 (Hable et al.)

Make a custom jittered kernel

Unique weights for r/g/b

Uncharted 2



Remaining Problems Much better performance, but still expensive

Doesn’t scale (10 heads/body-parts = 20 light-maps)

Seams in UV space

Light-map resolution

Banding due to inadequate samples


Screen Space SSS

Screen-Space Subsurface Scattering(Jiminez et al.)

Do the blur in screen-space

More scalable, but high fixed cost

Requires deferred renderer

Adds many G-buffer channels

Too many ‘S’s?


Pre-Integrated Skin Shading

Goal: A simple pixel shader All inputs stored locally in texture/bump maps (no blur passes)

Key Observation: Scattering isn’t visible everywhere

Occurs near changes in incident light (light gradients)

Strategy: Find light gradients and pre-integrate scattering


Constant incident light: Do nothing!


Small Surface Bumps


Shadows


Three Problems to Solve



1.) Surface Curvature

2.) Small Surface Bumps

3.) Shadows


Curvature: Pre-Integrated BRDFs

Diffuse Lambertian fall-off (dot(N,L)) Causes incident light gradient

Why not bake scattering in modified dot(N,L)? Wrap lighting (GPU Gems 1)

ATI Skin Fall-Off

Half-Life Fall-Off



Problems with wrap lighting? Not based on real skin diffusion profiles

Biggest problem: doesn’t account for curvature/thickness!



Problem: Not based on real skin diffusion profile Solution: Pre-blur diffuse BRDF with skin profile



Problem: Not based on real skin diffusion profile Solution: Pre-blur diffuse BRDF with skin profile

Problem: Doesn’t account for surface curvature Solution: Parameterize BRDF with curvature



How do we pre-integrate the diffuse BRDF?

Profile



Integrate all incoming light on a sphere(or a ring, it’s easier)

Profile



Let see look at some results!



Wrap Lighting GPU Gems 1 Gamma space lighting No tone-mapping


Curvature: Pre-Integrated BRDF

Wrap Lighting More recent (used today) Linear lighting Filming Tone-mapping



Wrap Lighting More recent (used today) Linear lighting Filming Tone-mapping

B/G/R Cutoffs

are clearly visible



Pre-integrated scattering Linear lighting Filmic tone-mapping



How does curvature help?



How does curvature help? Look at nose vs. forehead



Let’s decrease the size



Light scatters right through. Now let’s increase size



Flatter surfaces have less

visible scattering



Now bake everything!



So where do we get curvature (1/radius)



So where do we get curvature (1/radius) Similar triangles and derivatives.




c = length(fwidth(N)) / length(fwidth(pmm))




Derivatives are constant within a triangle Sometimes this is fine

If edges appear, store curvature in vertices or texture




Pre-integrated curvature-based BRDF

2.) Small Surface Bumps


Per-Channel Bent Normals

BRDF works for smooth curvy surfaces What about small bumps in normal maps?

Light should scatter past several surface bumps

Using curvature breaks down at small sizes

Let’s focus only on normals now



With a normal map we can sample neighbourhood We could take many samples, light, weight and accumulate

Or, can we pre-filter the normal map? Pre-filtering dot(N,L) is correct

Pre-filtering max(0,dot(N,L)) isn’t strictly correct (but it’s close)

LEAN/CLEAN mapping does this for specular



Tangent:

Normal maps can be acquired (Ma et al. 2007)

One technique applies light gradients to capture normals

Different light colours produced different captured normals



Ma et al. 2007 - Rapid Acquisition of Specular and Diffuse Normal Maps



We don’t need captured normals to use this technique Start with ‘true’ normal map (spec) Pre-filter new normal maps with R/G/B skin profiles Optimally we need 4 normals now

R/G/B and Spec

Looks great! Lots of memory!


Bent Normals (Optimized)

A.) Use geometry and spec normals (blend the rest)

Only works for certain art!

Normal map must contain only details (wrinkles/pores)

B.) Use two normal samples (blend the rest)

Works for all art!

One normal-map, two samplers

Clamp one sampler at blurred mip threshold


Bent Normals - Results

Without bent normals With bent normals


Bent Normals - Results

Now we have pretty good solutions for: Curvy but smooth surfaces (pre-integrated BRDF)

Flat but bumpy surfaces (pre-integrated normal maps)

These complement each other BRDF chosen from prevailing curvature (geometry)

Broad scattering from dominant surface

Bent normals fill in the gaps Fine-grained scattering of local details




Pre-integrated curvature-based BRDF

2.) ‘Small’ Surface Bumps

Pre-integrated bent normals

3.) Shadows

Pre-integrated shadow penumbrae


Shadows

What about shadows? Shadows also give skin a characteristic look Seems more difficult:

Shadows depend on arbitrary occluders

Can fall anywhere on the surface


Shadows

Turns out we can apply a very similar technique Other approaches have used a texture lookup to mimic

the look of skin(GDC 2005, Tatarchuk)

In our work we apply the diffusion profiles


Shadows

Turns out we can apply a very similar technique This is the shadow from a box filter

S(x)


Shadows


Given the shadow valueS(x)


Shadows


Given the shadow value

Invert the penumbra blur function

S(x)

S-1(x)


Shadows


Given the shadow value

Invert the penumbra blur function

Find the position within the shadow

S(x)

S-1(x)


Shadows

Map to a smaller shadow with room for scattering


Shadows

Apply similar integration as our diffuse


Shadows

Pre-integrate for different penumbra size (eg. due to slope)


Shadows

Let’s look at some results compared to TSD

Advances in Real-Time Rendering in GamesTSD

Pre-Integrated


Results

TSD


Results

Pre-Integrated


Results Diffuse


Results TSD


Results Pre-Integrated


Shadows

We need a wide penumbra to get lots of scattering We used two different shadow techniques

In both cases it really helps to use extra shadow maps We can cheat: only render the head, not the whole scene

Can be as low as 642 for one head (preferably 1282)


Shadows - PCF

1.) Pixel Quad Amortization (16 shadow map texels per sample!)

Pixels are rendered in quads – quad pixels can share work!

8x8 PCF in 4 samples!

Requires Nvidia card (but DX11 ‘fine’ ddx/ddy might fix that)

Tell me if you try it with DX11 hardware!!

See GPU Pro 2 for details


Shadows - VSMs

2.) Masked Variance Shadow maps Render only the head into it’s own small variance shadow map

‘Background’ causes bad variance artifacts!

Include an extra channel as background mask (eg. A2B10G10R10)

Blur the variance shadow map and mask

Correct variance sample to remove the background percentage

Correct the shadow calculation accordingly

Background percentage = not shadowed


Filmic Tone-mapping

All the images today use filmic tone-mapping Certain shots (eg. sunlight) need a lot of light Scattering occurs at the dark end of dynamic range

If we try capture the bright light

We invariably crush the darks into a small dynamic range

In summary, we need high dynamic range! And a way to map it nicely to a LDR display


Filmic Tone-mapping

Kodak film

response curves This is a nice

approximation: (Hejl 2009)

x = max(0, LinearColor-0.004);

GammaColor = (x*(6.2*x+0.5))/

(x*(6.2*x+1.7)+0.06);


Filmic Tone-mapping

Without filmic tonemap 1x intensity Larger size Boring!


Filmic Tone-mapping

Without filmic tonemap 4x intensity More interesting, but!

Saturated (clamped to 1.0)

Less saturated blacks


Filmic Tone-mapping

With filmic tonemap 4x intensity

Saturated blacks


Filmic Tone-mapping

Without filmic tonemap


Filmic Tone-mapping

With filmic tonemap


Filmic Tone-mapping

Digital cameras also have

this problem!

Left side is desired look Right side is blown out

Film is much better


Filmic Tone-mapping

Varying size

and intensity

by powers of two

Still 251/255



Future work: Approximate lookup textures with curve fitting

Use additional axis of curvature in diffuse fall-off

Include transmittance techniques

When the detailed normal map gets blurred, we lose the small scattering effects from our normal map technique. Possibly we should bias the curvature lookup to account for this in the BRDF.



Thanks to: George Borshukov

Natalya Tatarchuk

Google (where I work now)

Everyone at Electronic Arts (where I used to work)



Integrate all incoming light on a sphere(or a ring, it’s easier)

Profile Oops! saturate(cos())!

advances in real-time rendering in games. pre-integrated skin shading eric penner

Documents

realtime rendering

scattering slide

games curvature

skin profile slide

surface curvature slide

diffusion profile slide

realtime helps

realtime lots